Nice article Mike. By any chance, is there a possibility that the mobos might not be that efficient at handling either low or high power? Since there aren't really many other mobos to test along side with, that answer might be tough to answer but the system power draw looks like it could be lower for the mobo cpus. Might be forgetting something though...

Great article, makes me feel even luckier that my 3500+ has a TDP of 36.4w (although the TCaseMax is only 55c).

Since I've got it undervolted to 1.2v I'm guessing its about 25w at full load + CnQ sets it to 1ghz with 0.925v at idle

Isn't CnQ just going to set it back to the default vcore at 2.2ghz?

Not according to speedfan and XPC tools. I think the CnQ implementation on the shuttles is slightly different. It seems to just change the multipler between 11 (default) and 5, and drop / raise the vcore by ~0.3v. I would use Crystal but the bloody thing wont let me set the vcore below 1.1v. Its a shame because the CPU can do 2200@1.1v and 800@0.8v, but the machine requires a 1.2 vcore to cold boot.

Great work. Mike!
I'm now happier with my choice of A64 3000+

Btw, the first table in page 2 indicates (by mistake?) A64 3000+ a Winchester core, which is not consistent with rest of the article.

As fine a piece of work as I've seen in 45 years in computers.

Bravo.

Nice work guys!

One thing makes me a little concerned. My 6600GT, which I bought thinking it was a low-power card, actually draws almost double the amount of power in comparision to my undervolted Venice 3000+

The power from the 12V2 should be directed to the GPU instead of the CPU in todays systems!

I'm very pleased to see the Sempron 3400+ right up there in the power efficiency stakes; that confirms my own experience and exonerates my recent purchase of the same.



It depends what you define as 'small'. For example the Vcore difference between the highest and lowest P-states of my Sempron 3300 is only 0.3V, but at 1Ghz @ 1.1V it can be run passive, whereas at 2Ghz @ 1.4V it cannot. The relationship between CMOS supply voltage and power output is quadratic, so IMHO small reductions can make a big difference to overall heat output.

Don't the AMD CPUs have much more L1 cache?

And are there any plans to compare the power efficiency/usage of different motherboards (VRMs, but also chipsets)?

But you changed more than just vcore.

Yeah, L1 cache should be 128k for all the A64 chips.


I noticed a 6c drop in full load temps between 1.4v and 1.2v.

Mike,

While I appreciate the article. I can't help but wonder at some of the CPU choices. It seems like a lot of them were not the best representatives of their kind.

Was your Sempron a D0 chip?

It seems like it was because the voltage required for stable 2.0GHz matched my early 90nm D0 Sempron 2800+. My new E6 Sempron 64 2500+ was able to run the same speed at 1.08V (not to mention running at its default speed at 0.88V)

Pentium 4 630??? Why not the 65nm Pentium 4 631?

Core Duo T2600??? Why not the L2400?

Also it would have been nice if both the S754 and S939 motherboards at least used the same chipset. The power efficiency diferences between the different integrated chipsets would also be nice to know, but this review could have at least shown the difference single and dual channel DDR make using the same chipset.

Looking at the Sempron results seem to show that most of the Turion's advantage over the Athlon 64's is its platform, not the CPU iteself.

Also was no attempt made to undervolt the processors at the lowest state? It seems like the low state voltages of the A64's were way too high--seeing as my Sempron 2500+ can operate at 1.4GHz @ 0.88V. An E6 A64 should be able to go down to at least 0.85V @ 1.0GHz, if not lower.

I think there are several boards which operate like this. The Shuttle is probably just changing the actual voltage not manipulating the processor VID. Some boards can do both. I think my TForce 6100 operates like your Shuttle--at least for overclocking/overvolting it does. The effect of the C'n'Q changes is to manipulate the VID, the motherboard simply offsets the requested VID by the difference you've selected in the BIOS.

OK, I'll give you a different example. The P4-M chip in the computer that I am writing this on is currently running 1.86Ghz @ 1.038V; at exactly the same clock speed but increasing the Vcore to 1.3V, it can no longer be run passive (at least not without triggering thermal clock throttle mechanism); so we can see that the supply voltage to the CPU really is the primary determinant of heat output.

I think the point wasn't to test every possible CPU. But to give a general idea based on what CPUs they had available which are many of the more common ones.



Wow, 0.88V? I thought that most E6 cores had trouble getting below 1.2V I was excited that my E6 Sempron 64 2600+ got down to 1.05V. Although I didn't see much difference in thermals between 1.2V and 1.05V.

No, I just expect the lowest-power CPU's of each line tested. SPCR could sell off these old chips and buy these low-end/latest process processors. So, I can't see cost being the reason not to get the best examples.

This review just doesn't show the current state of the industry. There are just too few up-to-date processors for that.

Hardly, I even had a 130nm A64 2800+ running on 0.850V. The point is what clockspeed can these run that low on. For my $70 E6 S64 2500+ that is its default (rather low) clockspeed of 1.4GHz.

To agree with Jaganath, clockspeed is more of a linear relationship with power while voltage is quadratic (really fairly proportional to voltage squared). You may not notice a reduction in heat, but you will be realizing significant reduction in CPU power consumption. It is very probable that other components become the primary heatsources in the system.

Very kind words, thank you.


You must think we have big $$$ to spend on gear. Processors cost $, our finances are limited, and we take what we can get from whoever is willing to give/lend us the gear. The CPUs are a sampling of commonly used models. They were deliberately NOT chosen for min power, but for reasonably wide availability for average buyers.


rev DH8-E6 Venice. Re- your Sempron, I quote from the article:



Have you tried to buy one? Intel did not have a sample for me & it was not possible to find one in Vancouver in time for the testing. AFAIK it does not have wide distribution at this time.


Have you tried to buy one? See first comment above


Nice doesn't mean it's doable.

That's fodder for another article altogether, and this article was massive enough to complete. Besides, there are too many other factors (components on the boards) to isolate just the chipset alone. VRM efficiency & chipset both affect power readings, for example, as would different audio, graphics, LAN & other chips. This task would require test plaform choices of a different nature.

Wrong interpretation, imo.

Didn't you read the article? Each CPU was carefully undervolted to instability, then raised back up. Your experience differs from ours, but that's hardly proof of anything.

Finally, QuietOC, your comments suggest that you fail to appreciate the natural and unavoidable variances of computer/electronic gear (samples of the same model from the same batch are far from identical!!), the way different combinations of specific samples interact, and the difficulties involved in assembling so much gear altogether at the same time. If we'd detailed every product quirk. combination trouble and testing hurdle we went through to complete the article, it would be three times longer.

Mike, I did appreciate the article. I also greatly appreciate your site.

I am guessing that you can't sell all those old expensive chips then? The Pentium 631 has been available at Newegg for quite some time. And, yes, I do know that it is impossible to find new Intel chips through normal distributors.

FWIW: I have not observed large variations in the overclocking capability of CPUs of the same core/stepping. Perhaps I've always been very lucky. My experience has been mainly the low-end chips: Coppermine Celerons, Thoroughbred B's, and D0 and E6 Semprons/A64.


Yes, I see you found the lowest voltage for the high-state, but it looks like the low-state (the 1GHz speed) setting was using the default C'n'Q voltages. These are the voltages that affect the idle power consumption, and the greatest power savings would be relealized by finding the lowest voltages for that state.

Mmm, it's really unfortunate that Newegg does not ship to Canada.

Ah, I'm sorry, I misread your original comment. Yes, you are right, we did not check this. It could and probably should have been done, but with so much testing, it got missed. You could easily generalize, however, that a proportionate drop in idle power of at least 10% at the 2x12V socket could be seen across the board if the CPUs were undervolted at idle. If you charted this graphically, I doubt its shape would be much different than the stock idle power.

On the system power side, my guess is that the drop in power would be highest with the high power processors & virtually nil for the lowest -- the latter because VRM & PSU efficiency and residual power demand from other components would compress any power reductions. I'll see if any of the systems from the testing are still intact & if so, try to get readings on 3 processors -- a low, a mid and a high power one -- and work it into the article.

I remember P4 Northwood's with C1 stepping had very variable capability when it came to overclocking; factors such as location of manufacture, date of manufacture and batch number all seemed to be (contra-) indicators of overclockability. Some of the truly great C1 Northwoods (such as the SL6RZ and the SL6GQ) could hit 4Ghz and beyond, whilst the rubbish ones from the Philippines and Malaysia rarely broke 3Ghz.

However, recent CPU's from AMD do not seem to have this degree of variability in overclocking, although my experience with overclocking AMD chips can rightly be called minimal.

Great article. Answered most of my questions. Too bad I just got a sempron 3100+ for my next basic rig when i could have gotten the full A64 3000+ with the same power consumption. Oh well, the performance difference is probably only 5% and the A64 option is more money.

I'm not sure about all the people here on SPCR but i do know that a good number of them , including myself, fold 24/7. What is the power consumption during folding? I know it can't possibly be as high as cpu burn but would the full load numbers aproximate the folding numbers minus 10% or so? In this case our 'idle' power draw would be more similair to full load.

MikeC, wonderful job on publishing an article that involved so much work. I have been anxiously awaiting this article, but have to say it still leaves me with more questions.

1) Regarding thermals, there is no mention of CPU temps in the article. Is it safe to assume this is directly related to the CPU/system power consumption? (Noting that CPUs without the heat spreader - like the Turion probably cool better with their direct contact to the heatsink) If so, at what kind of wattage would one expect a difference in temps to start appearing in a fanned quiet rig?

2) Why is it that the different AMD CPUs have such different power consumption even when run at the same speed and voltages? Noted the sample variance and also the difference in TDP, but when we look at the idle draw with nearly all A64 CPUs at 1000mhz 1.18v, the 3500+ Venice stands out like a sore thumb with 10.4W. From your Tmax screenshots, it seems to have roughly the same TDP as the X2 4800+ which despite it's dualcore only consumes 6.6W or 63%.

3) I am a little disappointed that a closer matched comparison between the Turion and other A64 CPUs wasn't within the scope of the article to better answer the question if a Turion is really worth the problems to implement on a desktop (even though you did mention this would be a personal choice). The Turion idle was at 800mhz compared to the others 1000mhz. I know you mentioned you prefer a hardware/bios solution, but CrystalCPUID works very well and allows most A64s to go down to 800mhz without issue. Although probably not a big difference at idle, when we are talking only a few watts in the first place plus using 90% idle for oveall consumption, then it could be significant for comparison purposes.

I had really hoped a s939 Opteron would have been included with it's great undervolting capabilities. I know this is not a mainstream chip, but in my experience much more widely available than the Turion. I recently installed an Opteron 144 which Tmax rates at 44.1W TDP (Rev E4) on the same ASUS A8NVM-CSM and it happily ran Prime95 stable 48h+ at stock 1.8ghz undervolted to 1.0v (others seem to have similar experiences). CrystalCPUID allows the Opteron to be volted down to 0.8v, unlike the artificial 1.1v limitation it places on all other A64 rev E chips.

At idle, it comfortably runs at 1000mhz 0.8v and I suspect will nearly match Turion's consumption (considering similar TDP to the 4000+ San Diego, I measured approx 3.3W AC reduction between 1.2v and 0.8v) . Under load however the difference will probably be larger (with Turion ML's 35W TDP and MT's 25W TDP), however back to question 1), how many watts before there is really any significant temp difference? Under Prime95, I found reducing the vcore from 1.3v to 1.0v resulted in lowering AC draw by 9W. If we refer again to the similar TDP but higher clocked 4000+ San Diego, then I would expect the system draw to drop by at least the same 9W, bringing it within 13W AC of the Turion.


After all that work I'm sure you're just glad to have finished. But if you are willing to consider adding more CPUs to the list, I believe fellow SPCRers including myself would be willing to pay for the purchase of certain CPUs for testing and have SPCR send them to us after you've finished.

My E3 Athlon 3000+ can do about 1.125V at stock speeds just fine and probably far less with underclocking. I think there's plenty of overhead as long as you're not pushing the mhz too high. I currently run at 2.24ghz at 1.225v and I that gives me about a .05v cushion for stability over my lowest tested voltage. I've noticed some pretty big thermal differences myself, I can never get over 40-42C at full load with my current voltage/ambient temps and stock would definitely go around 50C.

Mike, if you would like to borrow my E6 Sempron 64 2500+, I'd be happy to send it to you. Let me know.

Mike, this is an awesome article and I love how much data is presented regarding the power consumption and large effects of undervolting and/or underclocking on power usage. I'm sure you spent far too much time on this and you have our thanks for doing so!

Personally I think there's a few things that could make this into one of the best articles ever made. The first thing would be having a few more examples of intel's mobile processors and at least one example of the lower power MT-series Turions. I'd love to know if there's any noticeable difference in the MT-series vs the ML-series when clocked at the same speeds/voltage. Since they weren't able to be incorporated into the original article due to a lack of the processors it seems, perhaps this would be a nice article to have updates on just like your recommendations lists. I'd love to be able to refer my friends to this and let them know their best options for low power when I help them build their computers.

Also I feel like efficiency is directly related to performance. None of us would obviously want a 2 watt 486 processor, even if it is the lowest power chip out of everything. The article is pretty good at matching up the amd processors against intel ones for performance purposes, however it doesn't really show a comparison/contrast in terms of what performance gains or losses would be tied into choosing a Turion 64 over an Athlon 64.

I know this could be found through any number of other hardware sites through digging through their benchmarks, but I think even presenting a few general benchmarks of performance and a "performance per watt" chart (with stock and undervolted results) compariing all the processors would make this article even that much more helpful for decision making. I also agree that for most articles, there's no reason to present performance information, but efficiency for CPU's is really tied into both how good the product is(performance mostly), and how much energy it uses. Not to mention, intel loves using performance per watt as a buzzword and it would be great to see how their current processors compared to what they're saying for merom/conroe.

Only one (petty) gripe: too bad you couldn't get the Venice 3000+ Socket 754 processor in the article. It's got a TDP of 51W (11W lower than the Sempron 3300+). I'm loving mine in my Media Center 2005 PC.

-Derek

Me too!


That's a corporate-political issue, not an SPCR issue IMHO. It does not belong in Mike's article.

If I hadn't made the mistake of posting a coupla trivially simple fan "review"s, I would not believe some of the above criticism! I independently discovered that SPCR persons are highly demanding; they want what they want and they want it now!

Yes; sometimes it's imperative to remember that SPCR basically operates on a shoestring budget, with only one full-time employee (AFAIAA). That said, the impulse to "improve" articles is a good instinct, as that way an article can "evolve" with later accretions making it into a kind of "super-article"; this CPU article is a good candidate for that process IMHO.